Getting Food Robots Out of the Lab — Rajat Bhageria

WEBVTT
00:00:07.732 --> 00:00:09.314
Welcome to the Build the Future podcast.
00:00:09.314 --> 00:00:11.727
My name is Cameron Wiese and I'm your host.
00:00:11.727 --> 00:00:29.251
I've always been fascinated by the ideas and the sentiment that drove American culture in the 1960s with the space race, a culture galvanized to dream about the possibilities of tomorrow, whether it's food, transportation, cities, biology or anything else.
00:00:29.251 --> 00:00:43.689
It was this cultural mindset, rooted in optimism, that the world tomorrow would be better than the world today, a mindset where people were compelled to build things and I quote JFK not because they were easy, but because they were hard.
00:00:43.689 --> 00:00:48.444
It's this desire to build and to dream that seems to have been lost.
00:00:48.444 --> 00:00:52.213
It's something we're here to bring back With Build the Future.
00:00:52.213 --> 00:00:58.865
We're here to promote the ideas and stories of those who see how the future can be better and promote their plans to get us there.
00:00:58.865 --> 00:01:09.227
It's our mission to get you to dream about the possibilities of tomorrow, to dream about the future that you want to live in, and inspire you to go build.
00:01:11.772 --> 00:01:13.174
Welcome back to the Build the Future podcast.
00:01:13.174 --> 00:01:19.132
Today we're talking with Brijat Bagheria, the founder and CEO of Chef Robotics Chef.
00:01:19.132 --> 00:01:29.754
They're not only doing AI and robotics stuff, but they actually have robots out in the market helping food manufacturers pack meals and meal kits you get through Joe's or Costco.
00:01:29.754 --> 00:01:35.998
They just come out of stealth and I'm excited to share this conversation with an adult and what the future holds.
00:01:35.998 --> 00:01:37.819
So with that, let's jump right in.
00:01:37.819 --> 00:01:39.221
Thank you for being here.
00:01:39.221 --> 00:01:43.090
I pulled Build the Future out of the light.
00:01:43.090 --> 00:01:44.542
It was like on the back burner.
00:01:44.542 --> 00:01:57.430
I'm like we got to jam because obviously there's a lot of interesting things going on, but I haven't seen anything on the robotics front that's actually like meaningfully deployed and like being used and growing.
00:01:57.430 --> 00:02:05.570
And then we got to catch them like okay, cool, we should, let's record something, because people don't think that robotics is actually happening.
00:02:05.570 --> 00:02:09.364
If they do, they're not exposed to it.
00:02:09.364 --> 00:02:16.307
So let's start with tell me about the future you're building with Chef Robotics and sort of the vision that you have for it.
00:02:17.020 --> 00:02:18.286
Yeah, and again, thanks for having me on.
00:02:18.286 --> 00:02:21.604
I'm super excited, I'm really glad we got connected and I'm excited to have this conversation.
00:02:21.604 --> 00:02:29.924
Yeah, so I think the future that we're trying to build is one where, you know, in the food industry, we find that these jobs are honestly very tough.
00:02:29.924 --> 00:02:47.394
A lot of the jobs are in either a very cold environment that's like a refrigerator, basically 34 degrees or it's a really hot environment and Americans basically just don't want to work in these jobs, and the reason is that you're just doing these redundant motions all day long and because of this, there's a really, really crushing labor shortage.
00:02:47.394 --> 00:02:53.203
So I think the future we'd like to build is one where you know, since you have robots doing these kind of redundant motions.
00:02:53.342 --> 00:03:01.207
So, for example, our initial use case is one where, you know, in these food factories, you essentially have these people who are in a 34 degree room.
00:03:01.207 --> 00:03:03.191
They're scooping food for eight hours a day.
00:03:03.191 --> 00:03:06.826
Oftentimes it's frozen material, so like their hands are numb.
00:03:06.826 --> 00:03:28.944
I mean, I tried this job for like five minutes and it's like my back is hurting and my arm is aching and like literally five minutes and, like you know, it's just a really tough job, and so I think we'd like to work a job where, like we like to create a future where we can have these kind of robots do that kind of day-to-day mundane work and now, with those other people, they can actually do something more better suited for people, right?
00:03:28.944 --> 00:03:34.526
So continuous improvement, line monitoring, robot operating things like that, and, of course, they're starting in one place.
00:03:34.526 --> 00:03:42.169
But really what I'd like to see is a world more generally where robots are doing these mundane tasks and humans are doing something better suited for humans, totally.
00:03:45.080 --> 00:03:45.983
Yeah, I think we'll sort of come back to that.
00:03:45.983 --> 00:03:49.276
But for the unacquainted like, what specifically are you guys doing at Chef?
00:03:50.221 --> 00:03:57.266
Within this kind of labor shortage in the food industry, we find that the majority of the jobs are actually in the food assembly portion.
00:03:57.266 --> 00:04:08.311
Food assembly is basically this idea of kind of taking pre-cut and pre-cooked ingredients and kind of portioning them at the right portion size and then placing them into whatever form you might want.
00:04:08.311 --> 00:04:14.872
That might be a prepared meal, that might be a sandwich, a wrap, it could be a pizza party tray, yogurt party tray, whatever you might want.
00:04:14.872 --> 00:04:19.576
So essentially, our goal is really to kind of deploy these robots.
00:04:19.576 --> 00:04:22.603
I know what we're doing right now is deploying these robots into food factories, right.
00:04:22.624 --> 00:04:35.911
So think about any kind of meal you might find on an airplane, or meals you might find at a hospital Even like Trader Joe's, right or Costco it's like, yeah, exactly Trader Joe's or meals you might find at any Starbucks, for example.
00:04:35.911 --> 00:04:40.507
That's kind of a starting point and it really makes a lot of sense because it's a really great use case for robots.
00:04:40.507 --> 00:04:41.447
It's pretty high volume.
00:04:41.447 --> 00:04:44.112
And then the idea, of course, is that scooping is scooping is scooping.
00:04:44.112 --> 00:04:51.889
If we can manipulate a chicken and do that consistently without damaging it, without crushing it, and we can do this for thousands of ingredients, then why stop there?
00:04:51.889 --> 00:04:53.122
Why not go to ghost kitchens?
00:04:53.122 --> 00:04:56.591
Why not go to fast casuals, prisons, hotels, universities, et cetera?
00:04:56.591 --> 00:05:06.843
Really, the ultimate goal is to build AI-enabled robots that kind of help automate the human part of this job that is really tough for people and then ultimately allow them to do other things.
00:05:07.524 --> 00:05:10.882
It was something that I hadn't really thought about until you and I were chatting initially.
00:05:10.882 --> 00:05:16.903
A lot of the food that you see at the grocery store is pre-assembled and it just didn't click.
00:05:16.903 --> 00:05:19.069
It wasn't something I'd consider.
00:05:19.069 --> 00:05:20.232
I'm like how does this get made?
00:05:20.232 --> 00:05:23.666
It's literally giant assembly lines.
00:05:23.666 --> 00:05:26.552
But everyone's like, oh, hasn't it been automated?
00:05:26.552 --> 00:05:28.002
No, it has not.
00:05:28.685 --> 00:05:38.845
Yeah, exactly, it's really fascinating actually because, like, even when we started Chef, we thought it was automated, which is why we our initial impression of where, to start with, actually fast, casual restaurants.
00:05:38.845 --> 00:05:44.305
But then we went to a lot of these factories and exactly what you said, it's these like giant assembly lines.
00:05:44.305 --> 00:05:49.545
You have 30 people on the line and they're scooping and we're like why isn't this done?
00:05:49.545 --> 00:05:51.040
Why isn't this done by automation?
00:05:51.040 --> 00:05:55.432
And what we learned is that our customers have actually tried traditional automation.
00:05:55.432 --> 00:05:57.663
It's just that traditional automation is very kind of.
00:05:57.663 --> 00:05:58.906
It's traditional for a reason.
00:05:58.906 --> 00:06:03.908
It's very simplistic automation, right, there's not many sensors.
00:06:03.908 --> 00:06:04.860
It's not really for a reason.
00:06:04.860 --> 00:06:08.262
It's very simplistic automation, right, there's not many sensors, it's not really making a decision, it's a bunch of stainless steel, basically, and maybe one motor.
00:06:08.262 --> 00:06:10.564
That's kind of just doing the same thing over and over again.
00:06:11.204 --> 00:06:12.283
So you can make it work.
00:06:12.283 --> 00:06:21.889
Let's say, for one ingredient, pick any ingredient you can make a dispenser work, but then if you change how you cut that ingredient or you cook that ingredient, then it won't work so well.
00:06:21.889 --> 00:06:23.190
Or you cook that ingredient, then it won't work so well.
00:06:23.190 --> 00:06:36.278
So you know, it's really good if you're, essentially if you have like five SKUs let's say I'm craft times right, I have five SKUs make for like ketchup, right, and five is an arbitrary number 10, 12, 15, whatever it might be A low number.
00:06:36.278 --> 00:06:43.863
Basically, you essentially get a dedicated custom line for each of the SKUs and you just run it all day long.
00:06:43.863 --> 00:06:54.755
On the other hand, if I'm making a meal, well, there's 200 different kinds of meals and so now you're not going to get 200 dedicated lines, you're going to get five to 10 flexible lines and you're going to change over.
00:06:54.755 --> 00:07:04.283
But of course, that flexibility is not something that depositors, suspensors, traditional automation can handle, which is why they end up using people and that's where we can come in.
00:07:04.303 --> 00:07:10.771
Yes, so deployment, walk us through sort of like what deployment looks like for you guys so right now it's humans on an assembly line, presumably.
00:07:10.771 --> 00:07:27.586
Food is sort of brought in, and then you have humans who are scooping whatever from the pre-assembled pan I think it's like the hotel style giant pans into these pre-assembled or the sort of preset containers, right, and then just one after another, over and over, and over and over again.
00:07:27.605 --> 00:07:30.322
Yes, exactly Exactly so, like a status quo is exactly that.
00:07:30.322 --> 00:07:33.762
So you have like each human right now has kind of a big tub and they're kind of scooping.
00:07:33.762 --> 00:07:41.365
Maybe the first two people on the line on either side are doing rice and the next person is doing a meat and the next person is doing some vegetables, et cetera, right?
00:07:41.365 --> 00:07:52.829
So the deployment process for Chef, we really designed it, engineered it, I guess I could say, to really be very simple, and the reason for this is each Chef we kind of call them modules, chef modules.
00:07:52.829 --> 00:07:55.629
They're the same footprint as a human, which is again on purpose.
00:07:55.629 --> 00:08:02.829
So the idea is that we're not automating like 10 people or 10 stations, if you will, it's automating one station at a time, right.
00:08:02.829 --> 00:08:09.187
So essentially the system itself is the same footprint as a human and it's on casters or wheels, right.
00:08:09.187 --> 00:08:20.250
So you literally slide it onto the line and then the only kind of two inputs we need is really a power line 110 AC, which every facility has, and then a compressed airline, which also every facility has.
00:08:20.250 --> 00:08:25.249
It's kind of like pneumatics, basically, and they use pneumatics for other equipment on the line, so that's why they have it.
00:08:25.249 --> 00:08:27.874
And basically you slide it onto the line.
00:08:27.874 --> 00:08:32.491
You kind of put down the forecasters and essentially you're kind of ready to run.
00:08:32.779 --> 00:08:38.383
At that point, what you're going to do is you're going to tell the robot okay, you know 6am, it's 6am time to run production.
00:08:38.383 --> 00:08:40.125
What meal am I making?
00:08:40.125 --> 00:08:41.768
Well, I'm making the pad time here, okay, great.
00:08:41.768 --> 00:08:46.453
Then it's going to ask you what ingredient you want to run and then you say, well, I'm doing the broccoli, Okay, great.
00:08:46.453 --> 00:08:49.827
Then it's going to say, great, well, you need to put in your trays of broccoli.
00:08:49.827 --> 00:09:11.826
You load the broccoli trays, hotel pans and the final kind of step is it's going to say, okay, well, based on the AI policy, i's like a dovetail mechanism to attach the utensil, there's two pins to hold it into place and essentially at that point, that's essentially the only kind of work that the line manager has to do.
00:09:12.327 --> 00:09:14.748
And at that point, basically, the robot will kind of fine tune.
00:09:14.748 --> 00:09:16.466
It'll kind of pick and dump is what we call it.
00:09:16.466 --> 00:09:23.730
You kind of pick ingredient broccoli from one tub to the other tub, over and over 10 times or so, until it kind of feels okay, I'm pretty consistent.
00:09:23.730 --> 00:09:27.509
At that point you press play and and that's kind of it.
00:09:27.509 --> 00:09:36.706
At that point the system will kind of see that there's trays going down the line and when it sees the first tray it'll say, okay, it's time to pick, and then it'll place and every tray it'll try to get better right.
00:09:36.706 --> 00:09:41.764
So we actually have weight scales underneath each of the pans to actually get some how much weight we picked up.
00:09:41.764 --> 00:09:45.169
Then we can say, well, it's too high or too low, let's actually improve the next time.
00:09:45.971 --> 00:09:46.933
Sweet, yeah.
00:09:46.933 --> 00:09:48.201
So I can imagine that like.
00:09:48.201 --> 00:09:57.731
Well, the interesting thing that I think happens here is, like most people, so you get it set up, but where does the more like technical component come in?
00:09:57.731 --> 00:10:02.702
Right Cause you have the stamp, you have the old, what this sort of automation name attempted to like pick up and go.
00:10:02.702 --> 00:10:06.115
Where's the insight?
00:10:06.865 --> 00:10:14.533
Yeah, I think the secret or the insight is that, dispensers, by their nature, you can't really have much sensing.
00:10:14.533 --> 00:10:16.192
It's kind of just doing the same thing over again, right?
00:10:16.192 --> 00:10:25.937
And I think our insight was okay, well, how does a human A human doesn't care if it's shredded chicken or sticky rice or cheese grits, they don't really care.
00:10:25.937 --> 00:10:27.587
Well, and how do they do that?
00:10:27.587 --> 00:10:31.049
Well, they use their eyes, of course, first of all to figure out okay, first of all, what ingredient is it?
00:10:31.049 --> 00:10:35.592
Second of all, okay, once I understand the ingredient, okay, well, how am I going to scoop?
00:10:35.592 --> 00:10:38.831
First of all, where in the hotel pan am I going to scoop, like the actual like where?
00:10:38.831 --> 00:10:44.332
But then also how right I going to have Right?
00:10:44.332 --> 00:10:47.879
So there's some, there's some perception and there's some intelligence there.
00:10:47.879 --> 00:10:49.942
And then, of course, they have arms through these motions.
00:10:49.942 --> 00:10:55.193
And the thing with scooping motions is that with a scoop, you can basically pick up anything.
00:10:55.193 --> 00:11:12.924
You don't care how it's cut, you don't care how it's cooked, you don't care how sticky it is, so long as you have the right place, you start in the right place and then you have the right motion, you can essentially pick up anything Right.
00:11:12.924 --> 00:11:13.943
And then the final portion is they have a utensil.
00:11:13.943 --> 00:11:16.852
And so we essentially said, okay, let's mimic that idea which is we're going to have computer vision to figure out where do I pick from and where do I place.
00:11:16.852 --> 00:11:26.394
We're going to have a lot of intelligence to have figured out what the motion is and also like, for example, like if I have a very sticky ingredient, I actually need a lot more force to actually kind of go through that material.
00:11:26.394 --> 00:11:31.604
Or if I'm picking broccoli, if I just like kind of dumbly go through it, then I'm going to like fling the broccoli off.
00:11:31.604 --> 00:11:39.292
So I need to actually kind of like delicately kind of scoop it in and then maybe jiggle a little bit so that if there's anything that's like about to fall out, it falls in.
00:11:39.292 --> 00:11:39.994
Things like that.
00:11:40.774 --> 00:11:46.538
And then utensils Basically it's kind of like computer vision, machine learning, motion planning and utensil design.
00:11:47.019 --> 00:12:30.470
So we essentially built our system around that idea which is scooping, and the idea was that every customer, no matter if it's Thai place or Indian place, it doesn't matter what the customer is, doesn't matter what kind of containers they have or conveyors they have or ingredients they have, so long as I have this module which has a six staff arm, an interface to have different utensils and the best CPU and GPU can get at the moment, then essentially I can mass manufacture that hardware, that module and ship it to every customer, no customization whatsoever for a customer, and then basically each ingredient can have its own AI policy and that might be like dozens of different parameters about how do I scoop that ingredient to be consistent, to not damage it, to not spill it, to get the right portion size, all these things.
00:12:31.371 --> 00:12:35.448
And finally the third part, of course, is the right utensils and the utensils.
00:12:35.448 --> 00:12:36.913
You have kind of a library of utensils.
00:12:36.913 --> 00:12:39.669
Once you have a good library, you can kind of mass manufacture them as well.
00:12:39.669 --> 00:12:43.207
You don't need to make something custom per customer after that.
00:12:43.207 --> 00:12:54.495
So I think that was the insight which is like, instead of kind of having this kind of hardware driven approach, which is very much kind of you know, do the same thing over and over, can we have more of a scooping or software driven approach which is a lot more AI driven?
00:12:55.186 --> 00:13:09.347
So tell me about how you guys are going about doing that, because a lot of the previous like AI, so like on the policies that mentioned policies, it's like okay, here's a predefined set of rules, right, predefined set of rules, right.
00:13:09.347 --> 00:13:12.441
When you drive like all the this is how wemo and cruise and tesla, or at least how they started out doing it which was like cool.
00:13:12.441 --> 00:13:14.827
If this, then that hundreds thousands, hundreds of thousands of times.
00:13:14.827 --> 00:13:22.409
But then something changed in probably what 20 like actually sort of like industry wise, like what?
00:13:22.409 --> 00:13:30.025
What change that doesn't enable sort of this like computer vision type approach yeah, no, it's a really, really good question.
00:13:30.066 --> 00:13:45.206
So, like you're right, like I think, since like 2014, even like basically what the software and car companies did and like what all the robotics companies did is essentially a rule based system and and rules by the way, there's still a lot of machine learning in the rules it's a rule.
00:13:45.206 --> 00:13:51.136
So, for maybe an example I can get from the AV world is like okay, I'm going a very simplistic example.
00:13:51.136 --> 00:13:54.429
I'm going to detect a red light and when I detect a red light, I'm going to stop.
00:13:54.429 --> 00:13:56.836
That is a rule right Now.
00:13:56.836 --> 00:14:03.772
The issue with that rules architecture, which is what you alluded to, is that you have a lot of like it gets very convoluted very fast, right.
00:14:03.772 --> 00:14:09.399
Okay, well, if you're a red light and you know you can, you can turn right, am I allowed to go or not?
00:14:09.399 --> 00:14:12.812
So, anyways, and that's a very simple, but there's thousands of these different rules you kind of create.
00:14:13.426 --> 00:14:16.668
We started Shaft the same way, right, which is OK.
00:14:16.668 --> 00:14:19.996
Well, essentially, we were like, ok, let's create like a config.
00:14:19.996 --> 00:14:21.990
That is kind of what we internally call it.
00:14:21.990 --> 00:14:26.035
So maybe, like sticky barbecue sauce is going to have a config.
00:14:26.035 --> 00:14:33.815
I just came up with that, but it's a good example, and the reason I came up with that is because it's sticky, okay.
00:14:33.815 --> 00:14:34.398
So what does stickiness mean?
00:14:34.398 --> 00:14:44.570
Well, if I am above the tray, once I've detected the tray that I want to place into and I've tracked it now I have to it's going to take, like, because it's sticky, it's going to physically take the material like a couple seconds to drop down.
00:14:44.570 --> 00:14:47.274
So we have this like config parameter dwell time.
00:14:47.274 --> 00:14:54.188
Dwell time is basically the time you like follow the tray before you move on, and that's a parameter that an applications engineer tunes.
00:14:54.830 --> 00:15:02.827
So an engineer basically creates that kind of backbone software and then an application engineer says, okay, because it's a very sticky ingredient, I need, I need to do a little bit more.
00:15:02.827 --> 00:15:13.442
Or maybe another example I can give is like we're talking about broccoli.
00:15:13.442 --> 00:15:16.499
So you know we have a utensil that's going to close around broccoli and you know, if I just dumbly close, I'm going to sometimes get one piece, sometimes I'm going to get three pieces and you're going to be very consistent.
00:15:16.499 --> 00:15:26.529
So we kind of create this like twist motion where basically we try to like fill the tool as much as we possibly can, but there's constraints because these are collaborative robots and if we apply too much force, then they'll protect to stop.
00:15:26.529 --> 00:15:27.791
It's a safety mechanism.
00:15:27.791 --> 00:15:30.712
It's meant to be collaborative.
00:15:30.712 --> 00:15:34.015
So there's a safety mechanism where if you put too much force it'll kind of stop.
00:15:34.015 --> 00:15:37.318
But of course, if you're trying to pick broccoli, it's very dense material.
00:15:37.318 --> 00:15:41.481
Yeah, it's more difficult than corn or zucchini or something yeah.
00:15:41.822 --> 00:15:42.062
Yes.
00:15:42.062 --> 00:16:00.047
So anyways, like and we can talk about this more but basically the way we actually got to production, basically there was a lot of deep learning and machine learning, but each ingredient had a rule about how to leverage that model to do something useful, and these configs had like 200 software parameters apiece.
00:16:00.047 --> 00:16:05.365
So that's how we got to market and what was beautiful about that approach is it allowed to ship?
00:16:05.365 --> 00:16:11.035
I think that was the most important, which is like what is the quickest path to getting something useful out in the world?
00:16:11.035 --> 00:16:14.047
And that's kind of what we started to ship.
00:16:14.047 --> 00:16:18.647
And what was nice about that is by shipping, we got real production data in the field.
00:16:18.647 --> 00:16:24.868
Right, we learned about okay, well, actually, the sauces actually are different densities day by day.
00:16:24.868 --> 00:16:32.653
Or if you have a pan of shredded chicken, it's actually different material properties as you go throughout the pan, because it's different parts of the bird, et cetera.
00:16:32.653 --> 00:16:35.354
So we learn these like things in the field.
00:16:35.354 --> 00:16:37.692
So now we get production data.
00:16:37.692 --> 00:16:45.356
We have a mapping from like image data, basically RGBD, to robot control data about what works and what doesn't work, robot action data.
00:16:45.356 --> 00:16:48.775
How do we actually, in production, be consistent?
00:16:48.775 --> 00:16:49.548
How do we not spill?
00:16:49.548 --> 00:16:50.975
How do we not damage material.
00:16:50.975 --> 00:16:55.812
How do we place in the right compartment without spreading or spreading based on what the customer wants?
00:16:55.812 --> 00:17:06.571
And now that we have that production data, now we've started more of an end-to-end approach, which is kind of say, okay, let's go from images and input data to more core robot control data.
00:17:06.952 --> 00:17:11.059
But I think that was only possible because we started with production data.
00:17:11.059 --> 00:17:21.454
And the reason I say this is, when it comes to manipulation, there's no deformable physics simulators that really exist that are really good.
00:17:21.454 --> 00:17:26.051
So what a lot of companies do, even autonomous vehicle companies, is that they can train in simulation.
00:17:26.051 --> 00:17:31.182
Well, chef can't really train in simulation because there's no physics simulators for deformable materials.
00:17:31.182 --> 00:17:34.732
So the only way for us to get that data was going to production.
00:17:34.732 --> 00:17:49.971
So, anyways, that was kind of the story right, which is like you have something useful in the world by using this kind of approach that software running driving cars and others came up with, which is kind of like you have rules based on different kind of parameters and then you ship that, you get production data and once you have production data, now you can do a moment and end it.
00:17:50.465 --> 00:18:00.891
Well, and like, I think that's exactly sort of why what I was like fishing for Because the thing that I just the physical world's so complicated and there's all this.
00:18:00.891 --> 00:18:11.752
I am a hundred percent like profuge, like let's build stuff, but there's this sort of narrative around AI and robotics and oh, it's going to take over everything, and it's like it's not that simple.
00:18:11.752 --> 00:18:14.916
These things are really complicated and there's no training data.
00:18:14.916 --> 00:18:17.079
There's like, okay, I'm going to pick up this cup.
00:18:17.079 --> 00:18:21.273
Okay, well, how do you know?
00:18:21.273 --> 00:18:22.605
This is paper versus ceramic versus like metal.
00:18:22.986 --> 00:18:30.275
And then it just, and it's even deeper than that, right, just, and, and it's even deeper than that, right, but like from the outside, so it's even deeper.
00:18:30.275 --> 00:18:32.061
So, like from the outside, you can't tell how much coffee is inside that cup.
00:18:32.061 --> 00:18:36.153
So, and and, and, by the way, like a human, like, in other words, what I'm saying is vision is not enough.
00:18:36.153 --> 00:18:40.136
So what a human does is that they use their vision to figure out okay, generally, where do I pick from?
00:18:40.136 --> 00:18:48.332
But then they there's like appropriate reception, right, so they actually figure out oh, let me, how much force do I need to impart so that I don't that cup doesn't fall out?
00:18:48.332 --> 00:18:49.472
That's all.
00:18:49.472 --> 00:18:52.736
That's all real time that humans are doing that, but it's really hard for robots.
00:18:52.736 --> 00:18:56.038
No, so your point is exactly right, which is that I think.
00:18:56.460 --> 00:19:09.576
I think the sense I get is that and this is just a sense, my opinion right Is that I think a lot of people looked at what happened with large language models and ChachiPT and they're like, what used to happen in the world is we used to have these point models.
00:19:09.576 --> 00:19:10.699
You have a model that plays chess.
00:19:10.699 --> 00:19:12.060
You have a model that drives a car.
00:19:12.060 --> 00:19:16.295
You have a model that, like you know, does some natural language processing, so to say.
00:19:16.295 --> 00:19:20.156
And now can we have this kind of foundational model where it can do anything?
00:19:20.156 --> 00:19:32.519
And I think that paradigm actually works quite well for kind of language right, and the reason it does well for language and kind of written stuff basically code and things like that or purely the digital stuff like an image event.
00:19:32.519 --> 00:19:33.666
It's a purely digital idea.
00:19:33.666 --> 00:19:41.830
It can work well for that, because there's a lot of kind of language image data pairs on the internet that you just download.
00:19:41.830 --> 00:19:50.958
So I mean all the LLMs and GPTs and all these foundational models are essentially trained on kind of more or less off the shelf data from the web.
00:19:50.958 --> 00:19:55.952
And now of course, more people are getting proprietary data with proprietary sources, but generally it's kind of off the shelf.
00:19:56.065 --> 00:19:59.490
Well, I think with the robotics it won't just be like a snap.
00:19:59.490 --> 00:20:01.611
It was like I think Chachi PD was kind of like a snap right.
00:20:01.611 --> 00:20:02.615
It felt like it, if you will.
00:20:02.615 --> 00:20:13.842
I mean, obviously they've been working on it for years, but for the outside world it was like we suddenly have this kind of relevation of transformers and things like that.
00:20:13.842 --> 00:20:22.865
I think that's not necessarily going to be true in robotics, because the big missing thing is exactly what you said, which is that training data and I would even kind of go so far as to say that a lot of that the simulation is good.
00:20:22.865 --> 00:20:31.992
But even for non-deformable let's say I was working on a bin picking company or I was working on a palletizing robot company Even then sim is kind of not sufficient, I think.
00:20:32.545 --> 00:20:36.374
I think you really just and the reason for this is the physical world is very highly dimensional.
00:20:36.374 --> 00:20:39.865
There's a lot of noise in the physical world, like in simulations, everything kind of what you expect.
00:20:39.865 --> 00:20:50.338
But in the physical world there's like I mean, in the palletizing case, there's like forklift drivers that are coming in and out of the workspace, there's people who are like doing dumb things, there's weird kind of orientations of the pallets.
00:20:50.338 --> 00:20:51.641
So there's all this noise.
00:20:51.641 --> 00:20:56.827
You can try to simulate it, but it's always going to be limited compared to what's going to happen in the physical world.
00:20:56.827 --> 00:21:02.727
And I think the thing then becomes what is the fastest pathway to getting data from the physical world?
00:21:02.727 --> 00:21:05.051
And I think the fastest pathway is shipping robots.
00:21:05.051 --> 00:21:08.818
But now you have this chicken and egg problem, right, how do you ship robots without that?
00:21:08.818 --> 00:21:18.410
So I think it'll happen, but I think it's just going to require companies that are like really maniacally focused on shipping into production, getting data and then using that to make their models better.
00:21:19.406 --> 00:21:23.807
This is kind of the argument for general purpose Robots.
00:21:23.807 --> 00:21:31.930
Right, like you're, like Rosie from the Jetsons River, you know I love optimists aesthetically and figure cool.
00:21:31.930 --> 00:21:41.256
I just have a hard time seeing, like, where, like, how do they design these things so that they they can like learn to, like move your laundry?
00:21:41.256 --> 00:21:44.057
It's like it's so complicated, like, how do you?
00:21:44.057 --> 00:21:46.898
I don't know, this is a little bit of a deviation, but how do you think about this?
00:21:47.799 --> 00:21:48.840
I 100% agree.
00:21:48.840 --> 00:21:50.221
I mean, you know one.
00:21:50.221 --> 00:21:52.282
There's like a bunch of thoughts about this.
00:21:52.282 --> 00:21:59.107
So, like you know, one way, like a lot of people are comparing like these general purpose robots to self-driving cars.
00:21:59.107 --> 00:22:11.186
Well, the reality is that self-driving cars, like the first DARPA Grand Challenge that really in my opinion and others might again disagree with this, but I think AV in my head kind of like kicked off with the droid bed band challenge where, like Sebastian Thread and some of these other folks kind of like first competed.
00:22:11.186 --> 00:22:12.790
That was like 20 or so years ago.
00:22:12.790 --> 00:22:20.413
So from the time that competition happened right in the early 2000s to now was two decades ago.
00:22:20.973 --> 00:22:27.653
And I think my opinion is that general purpose robotics and humanoids will probably take around the same.
00:22:27.653 --> 00:22:34.986
And one way, it's basically a time horizon question, right, it's like maybe that happens, but how long is it going to take?
00:22:34.986 --> 00:22:38.114
And then there's some other kind of angles to think about it from.
00:22:38.114 --> 00:22:42.654
So one angle to think about it is like essentially the walking part is honestly not the hard part.
00:22:42.654 --> 00:22:43.928
I think we've kind of figured out walking.
00:22:44.368 --> 00:22:49.607
The hard part is manipulation, right, and yet we have not as a humanity figured out manipulation.
00:22:49.607 --> 00:22:51.355
I mean like forget deformable goods.
00:22:51.355 --> 00:22:55.734
What we're doing is deformable, like food is hard in and of itself, but we haven't even figured out non deformable goods.
00:22:55.734 --> 00:23:04.309
I mean, there's companies, many companies, and we can talk about some of them in particular if you'd like, but bin picking is just not a solved problem by any stretch of the imagination.
00:23:04.309 --> 00:23:11.901
So now we still haven't solved that basic like how do you pick up that like coffee cup you just talked about, or different objects you might have around?
00:23:11.901 --> 00:23:13.607
We haven't even figured that.
00:23:13.907 --> 00:23:25.409
And now we have to add that on to this other problem, which is okay, well, half of my compute, half of my load resources are being used to like not fall down and like there's a lot of safety ramifications.
00:23:25.409 --> 00:23:27.251
Anyways, all these will happen.
00:23:27.251 --> 00:23:34.750
I'm not saying that won't happen, I just think that like it's a little bit early, which is to say like I think we need to just solve manipulation, adding legs onto it.
00:23:34.750 --> 00:23:49.990
It's a good vision, right, but it just we just need to solve the manipulation part and I think, just focusing on just getting manipulation in a good place, then you can have another like put instead of an arm mounted like for chef, instead of an arm mounted on like casters, you can put an arm on on legs.
00:23:49.990 --> 00:23:51.518
That's fine, it's great.
00:23:51.518 --> 00:23:55.351
But just the hard part is manipulation, and just focusing on the problem is an important thing.
00:23:55.632 --> 00:24:04.148
You can't just, you can't just toss the legs on it right now and be like oh cool, now it's somewhere where it's like yeah, so it's more of an order of ops, yeah it just exactly, and this is so easy to do.
00:24:04.229 --> 00:24:06.692
Demos, I mean, I think there's a lot of really compelling demos.
00:24:06.692 --> 00:24:16.594
I think what I'm really impressed about is when somebody actually ships production robots, because in production, you know, manufacturers just don't care, they're just like output, right.
00:24:16.594 --> 00:24:21.846
They don't care about the technology, they don't care about imitation learning, they don't care about transformers, they just care about output.
00:24:21.846 --> 00:24:26.074
Give me reliability, give me throughput, give me flexibility, give me the things they care about right.
00:24:26.074 --> 00:24:29.058
I'm really impressed by the companies that actually ship in production.
00:24:29.500 --> 00:24:31.932
Yeah, and you guys are one of the leaders.
00:24:31.932 --> 00:24:35.536
How many meals have you guys helped assemble?
00:24:36.185 --> 00:24:37.388
Right now we're at 22 million.
00:24:37.388 --> 00:24:39.476
I think 22 million Incredible.
00:24:40.005 --> 00:24:49.656
So you're getting real data and real customers are using this to like improve their, their processes, which I think is just like again.
00:24:49.656 --> 00:24:54.575
That's why I'm like, okay, we gotta, we gotta do a pod, because it's like you don't hear that every day.
00:24:54.575 --> 00:25:00.675
It's like oh, there's another demo, one on x, or you know like someone got a funding round for this.
00:25:00.675 --> 00:25:02.506
Like do you have stuff in the market?
00:25:02.506 --> 00:25:04.751
Like so, like who.
00:25:04.751 --> 00:25:08.248
And I think back to your question like who is who else has stuff in the market.
00:25:08.248 --> 00:25:09.070
You're like, okay, this is this question.
00:25:09.070 --> 00:25:10.096
Like who is who else has stuff in the market?
00:25:10.096 --> 00:25:12.045
You're like, okay, this is, this is cool, like this is meaningful, like there's stuff going.
00:25:12.805 --> 00:25:18.996
I mean, there's a few that I'm like I'm really impressed by, and this first one I'm going to say is maybe like one company that actually has done it well.
00:25:18.996 --> 00:25:19.886
So there's this graph.
00:25:19.886 --> 00:25:28.038
I saw, and maybe you've seen it, like the Tesla FSD miles driven right, right, and I don't know if you've seen it's like kind of exponential curve and they're doing like a billion miles.
00:25:28.038 --> 00:25:29.819
I think that that.
00:25:29.819 --> 00:25:34.762
So I I think here's what's funny, right and and again, like there's not like a right answer for this, just different opinions.
00:25:34.762 --> 00:25:45.354
I remember, like when I first started in robotics, like everyone was essentially making fun of tesla for their whole approach of like you know, obviously cameras, but also other things, but just shipping something.
00:25:45.354 --> 00:25:57.349
And a lot of people had this approach of, okay, no, like we're going to raise tens of billions of dollars in funding, like the Waymo approach or the cruise approach, right, and we're going to try to learn, so to say, in a lab, and then over time we'll get to a full autonomy.
00:25:57.349 --> 00:26:00.787
And Tesla said, no, we're going to ship a bunch of vehicles, get production data and improve over time.
00:26:00.787 --> 00:26:13.301
I'm not saying that that approach has definitively won, so to say, but it really seems like people are pretty happy with FSD 12, the imitation learning part of it, and it seems like that's working and that vision is exciting.
00:26:13.301 --> 00:26:23.411
So if you look at like Tesla as a robotics company, arguably they're an extremely successful robotics company because they have production robots that are using imitation learning to actually like, successfully drive around cities.
00:26:23.411 --> 00:26:24.214
That's pretty cool.
00:26:24.214 --> 00:26:29.550
Other robotics companies I'm really impressed with probably the like and this one's a little bit like.
00:26:29.550 --> 00:26:33.959
It's an interesting example, but I think it's still it's necessary to call out which is Amazon Robotics.
00:26:33.959 --> 00:26:40.290
So Amazon Robotics actually has they have the biggest fleet of robots in the world and I say intelligent.
00:26:40.290 --> 00:26:44.154
Now kind of the ding again and that's hundreds of thousands of robots.
00:26:44.154 --> 00:26:48.230
Right, the ding against Amazon is that, well, it's not the ding, it's still very impressive.
00:26:48.230 --> 00:26:51.096
It's just like they're not deploying these robots at other customer sites.
00:26:51.096 --> 00:26:58.518
So if you're deploying your own facility, you can kind of engineer the facility to be perfect for robots.
00:26:58.518 --> 00:27:05.228
Let's make it really easy, let's have April tags everywhere, let's really basically just like, make it perfectly amenable for robots.
00:27:05.228 --> 00:27:16.855
But obviously, if you go to a different customer site which is what we have to do and other robotics companies have to do then you kind of have to deal with their randomness, that high dimensionality, which you don't have to do if you'd engineered them in space.
00:27:16.855 --> 00:27:20.136
But anyways, I think it's really impressive nonetheless, what they've been able to do.
00:27:21.005 --> 00:27:22.767
A couple others that I'll call out that I think are impressive.
00:27:22.767 --> 00:27:27.212
So I think the colloquial one that everyone talks not everyone, but many people talk about is locus robotics.
00:27:27.212 --> 00:27:30.717
So locus robotics makes these kind of amrs.
00:27:30.717 --> 00:27:36.271
These are robots that kind of help with like picking, you know, in a warehouse, a 3pl or e-commerce warehouse.
00:27:36.271 --> 00:27:41.909
Right, you need to like assemble a, a kit or a, you know, basically move materials around.
00:27:41.909 --> 00:27:49.286
And you have a human who's kind of like taking an object here, taking an object here, taking an object here, and these robots essentially kind of just help them do that and kind of follow.
00:27:49.286 --> 00:27:50.949
It's not actually doing the manipulation.
00:27:50.949 --> 00:27:53.652
The human is actually doing the manipulation, but the robot is actually kind of just like.
00:27:53.652 --> 00:27:56.075
Essentially it's a guy if you think about it.
00:27:56.075 --> 00:28:00.181
It's kind of like a moving shopping cart, if I were to simplify it quite a bit but it'll also do other things.
00:28:00.181 --> 00:28:03.709
They move pallets around and stuff like that.
00:28:03.709 --> 00:28:05.211
So they've deployed a lessons to be learned from them.
00:28:05.271 --> 00:28:13.076
But I think the point still stands, which is there's not hundreds of robotics companies who have shipped production robots.
00:28:13.096 --> 00:28:14.428
There's like tens that are.
00:28:15.270 --> 00:28:18.000
And when I say production robots, I mean intelligent robots.
00:28:18.000 --> 00:28:25.711
There's like thousands of companies that have shipped dumb robots, which are these kind of robots you might find in car factories that are doing the same thing over and over again.
00:28:25.711 --> 00:28:30.453
Right, these are robots that are essentially hard-coded.
00:28:30.453 --> 00:28:32.991
They've been around for 40, 50 years, kind of a solid problem.
00:28:32.991 --> 00:28:40.571
You go to Systems Integrated, they'll make you, they'll buy a robot from FANUC and they'll like, do some PLC programming and like and do the same thing over and over again.
00:28:40.571 --> 00:28:46.571
That's kind of a again, I don't want to say solid, but I think it's a pretty stable place.
00:28:46.571 --> 00:28:58.317
I mean, I think you can basically solve any problem you want within some limitations, but essentially most problems you can solve that way for applications where it's very low flexibility needed.
00:28:58.317 --> 00:29:04.971
So if you just want to do the same thing over and again, like I'm making millions of iPhones, done right, just get a custom line that just does that.
00:29:04.971 --> 00:29:11.297
But when it comes to intelligent robots that are kind of making decisions and having to be flexible, there hasn't been a ton.
00:29:11.738 --> 00:29:16.525
Yeah, I think that's a really, really important distinction for people, because they're like oh wait, they have robots in factories.
00:29:16.525 --> 00:29:19.454
It's like, yeah, for giant assembly lines.
00:29:19.765 --> 00:29:38.643
But the sort of places where there's a little bit of like, at least right now, human intervention needed is the place where, like, that hasn't been entirely taken over yeah, I mean a good example just to like finish the car analogy is, like you know, the the first kind of kind of parts, the early parts of car manufacturing line will will actually be fairly automated.
00:29:38.663 --> 00:29:49.693
There's these giant fanuc and kuka and abb arms if you were to go to the ford plant or the tesla plant, really any any kind of car oem but then you go to general assembly, which is kind of the last phase and it's extremely complex.
00:29:49.693 --> 00:29:56.094
And I mean the classic example a lot of people talk about is like kind of adding the wire harness right, it's like this kind of deformable, bendable thing.
00:29:56.094 --> 00:29:59.791
You have to like finagle this around the chassis of the car.
00:29:59.791 --> 00:30:02.506
I mean that's just very hard and you have all these people doing it.
00:30:02.506 --> 00:30:04.328
But but there's a lot of examples for that.
00:30:04.328 --> 00:30:10.700
Where the last phase is very flexible, there's a lot of these like deformable parts or moving parts and changes.
00:30:10.700 --> 00:30:14.971
So that's a lot of humans and that's where intelligent robots would be in this, in this equation.
00:30:15.605 --> 00:30:22.827
So it's like the problem like last mile delivery, it's like it's really really good up until that, also bet, and then it gets super complicated.
00:30:22.827 --> 00:30:40.628
Okay, I think that's a really good sort of picture of the landscape on robotics right now that most people, myself included honestly like may not be completely aware of, like is there anything else you think the average like person or general public like may not know about robotics or that may have like misperceptions about?
00:30:40.849 --> 00:30:41.873
misconceptions about right now.
00:30:41.873 --> 00:30:43.336
It's a very good question.
00:30:43.336 --> 00:30:47.830
So, like I mean, I think like, yeah, we kind of talked about the data bit, which I think is the really hard part.
00:30:47.830 --> 00:30:58.212
We kind of talked about there are these production robots, but they're for low mix applications and high mix, so to say, where you have varieties, where you need intelligent robots, and we talked about how it's different from like language.
00:30:58.212 --> 00:30:59.971
I guess a few other things I'd call out.
00:30:59.971 --> 00:31:03.309
I think you know here's actually an interesting one.
00:31:03.349 --> 00:31:15.069
So I think robotics is one of those industries more like in, you know, in startup world we often hear about, like in polygram talks a lot about this, but but I think a lot of people talk about how the number one reason startups fail is they build something nobody wants, right.
00:31:15.069 --> 00:31:17.737
I think that's especially true in robotics.
00:31:17.737 --> 00:31:19.351
I would say, and why is that is?
00:31:19.351 --> 00:31:24.829
It's because, like in robotics, it's these very technical founders, right, like these really incredible engineers.
00:31:24.829 --> 00:31:28.829
It's very easy to kind of build something that's kind of a cool technology.
00:31:28.829 --> 00:31:36.036
It's very easy to build, but it's really hard to take that next step with building a product and it's even harder to take that further next step which is a solution.
00:31:36.036 --> 00:31:38.411
So here's what I mean by that, you know, in the food.
00:31:38.411 --> 00:31:39.453
I'll just talk about the food world.
00:31:39.453 --> 00:31:48.759
I mean in the food world, like you cook at home a lot maybe not you, but I'm saying people cook at home a lot and they're like, oh, cooking is what I should automate.
00:31:48.759 --> 00:31:52.211
So you have these like robots that do cooking, like burger flipping, things like that.
00:31:52.211 --> 00:32:02.674
The issue, you find, is that if you actually talk to customers and you actually really understand what they're facing, what you find is that cooking actually scales relatively, not sublinearly.
00:32:02.674 --> 00:32:17.817
You know it takes one person to cook at your house for a few people, but it also takes one person to flip 50 burger patties, like one person can actually do a lot, because one person can kind of it's kind of batches basically, and so the pitch of robotics is really tough.
00:32:17.817 --> 00:32:20.946
For example, that's something that's not intuitive.
00:32:20.987 --> 00:32:34.223
And if you were to just build a technology, you end up building a cool cooking robot, but that actually really does not solve any customer problems, because you're going to go to the customer and say, hey, we have this cool cooking robot, but that actually really does not solve any customer problems, because you're going to go to the customer and say, hey, we have this cool cooking robot and they're going to say, okay, great, how many people can I kind of offset and have them do something else?
00:32:34.223 --> 00:32:40.566
And you're going to say, well, you can offset them, but our robot's going to fail sometimes you still need that person actually.
00:32:40.566 --> 00:32:46.353
And they're going to be like, well, there's not really an ROI anymore, right, right, so that that's something that's like not obvious, right, I think.
00:32:46.353 --> 00:33:00.065
So One of the things we try to do is actually sell the product before building it, and what that gets you is like you really like as soon as you have to ask for money, people really tell you what they really want and you can actually work backwards from requirements, as opposed to let's just build a cool thing.
00:33:00.065 --> 00:33:04.970
That same idea, I think, is true in a different way.
00:33:04.970 --> 00:33:23.834
So, like now, let's say, you have a general kind of thing that people want, right, when it comes to robots, I think even then it's kind of really easy to focus just on the core, like computer vision and motion planning or AI or whatever hardware, whatever it might be, and sometimes not think so much about the rest of the solution.
00:33:23.834 --> 00:33:24.655
So here's what I mean.
00:33:24.997 --> 00:33:37.526
So in Chef world, we got to make something that's very flexible and the more ingredients it can do, the more like different kinds of, you know, consistencies it can deal with, and textures and things like like how do we pick up a blueberry without squishing it the more we can do.
00:33:37.526 --> 00:33:38.568
Obviously, it's very useful.
00:33:38.568 --> 00:33:41.336
It's very, very useful Now, but it's not so.
00:33:41.336 --> 00:33:41.664
It's not.
00:33:41.664 --> 00:33:43.067
It's necessary, but it's not sufficient.
00:33:43.067 --> 00:33:51.881
So what also matters is things like okay, well, the users who are going to use this line are usually non-English speaking, often not college educated.
00:33:51.881 --> 00:33:52.823
They're not.
00:33:52.823 --> 00:33:56.353
They haven't really dealt with technology outside of maybe their like smartphone.
00:33:56.353 --> 00:33:58.079
So, okay, like it's.
00:33:58.099 --> 00:34:01.671
Things like the things that matter also like okay, is the entire interface in Spanish?
00:34:01.671 --> 00:34:05.567
Is the entire interface as simple to use as, like Netflix?
00:34:05.567 --> 00:34:09.152
The things that also matter might be like how easy is it to clean the thing?
00:34:09.152 --> 00:34:11.916
Like sanitation, how human, safe is it?
00:34:11.916 --> 00:34:14.364
I mean, these are things that are honestly not sexy, right?
00:34:14.364 --> 00:34:26.329
Like, making an interface into Spanish is not the most technically complex thing you might have to do, but that's what separates a technology into a product and a product into a solution, because I can have the most flexible robot.
00:34:26.329 --> 00:34:32.800
By the way, maybe somebody can build a robot that's extraordinarily technically competent, but if it's not usable.
00:34:32.800 --> 00:34:34.090
People don't want to use it on the line.
00:34:34.090 --> 00:34:38.153
Or the sanitation team hates to use it, or the production team finds it annoying.
00:34:38.153 --> 00:34:39.909
They're not going to use it.
00:34:40.885 --> 00:34:45.110
And so you have to sort of feel like it's all sort of like integration into again into the real world.
00:34:45.110 --> 00:34:49.195
It's like, hey, how's this going to interface with everyday people?
00:34:49.195 --> 00:34:50.552
Right, it's not like.
00:34:50.552 --> 00:34:55.990
I mean, there's sort of an argument for like on the ghost kitchen stuff, like if you were fully automated.
00:34:55.990 --> 00:35:02.612
Like there's this restaurant in San Francisco that had this sort of like assembly line where I was just picking the ingredients.
00:35:02.612 --> 00:35:13.378
Like you had a whole team of people who were at some industrial kitchen who are prepping all the ingredients and like loading them in, and so it's like it ends up being kind of at least in that it's like kind of gimmicky.
00:35:13.378 --> 00:35:17.492
Same with the, the robot cafe x, yeah, like sfo.
00:35:17.492 --> 00:35:22.367
It's like it's cool oh, I think photos of it, but it's not actually solving a problem.
00:35:22.367 --> 00:35:37.291
Like I'd rather go to the starbucks so I can talk to the barista and be like hey, actually can I do this without like six pumps of like vanilla syrup, because I don't want nobody who wants that or I don't want that versus like entertainment.
00:35:37.351 --> 00:35:40.125
I think it's interesting to sort of differentiate between like with robotics.
00:35:40.125 --> 00:35:42.349
It's like what's what's being done for entertainment?
00:35:42.349 --> 00:35:44.514
Like battle boss, great cool.
00:35:44.514 --> 00:35:48.211
Like more of that please, but larger, ideally.
00:35:48.211 --> 00:35:51.246
Yeah, they're not actually solving a problem Exactly Like if we were to.
00:35:51.246 --> 00:36:03.197
If we sort of like imagine, imagine a future where robots are solving like a lot more like real problems, not just superficial ones like what, what does that world look like?
00:36:03.197 --> 00:36:05.373
And so this sort of connects to you, mentioned to me.
00:36:05.373 --> 00:36:17.235
You're hoping that with Chef, you inspire a lot of other people to do more like real, practical, impactful things with AI and robotics, and I want to sort of see if I can get you up on a soapbox to riff on what some of those things might be.
00:36:17.784 --> 00:36:18.989
Yeah, I mean it's really important.
00:36:18.989 --> 00:36:22.869
So, like actually early on in the Chef history, like again, robotics is not easy, honestly.
00:36:22.869 --> 00:36:29.807
So there was a moment in the early days when I was like this sucks right.
00:36:29.807 --> 00:36:33.791
Like you know, like it's like every day is hard, like dealing with physics is hard, everything's just hard about it, right.
00:36:33.791 --> 00:36:43.753
And then you, even even if you get the thing to work, then the customer is like, oh, but this isn't, it's just like everything's hard and there's supply chain issues and there's deployment issues and customer support issues, and customer support is just not on the online.
00:36:43.753 --> 00:36:50.869
It's like you go to the customer site sometimes, etc.
00:36:50.869 --> 00:36:55.144
Anyways, there's a moment where I like I had a conversation with another friend in the software chuck space and I was like this sucks, like why are we doing it?
00:36:55.144 --> 00:36:57.413
Why don't we just do like sass or something, just something easier?
00:36:57.413 --> 00:37:05.951
Like, basically, I was considering quitting, right, and he had a really interesting response which is like I still remember to this day, which is like, yes, it's really hard, but that's kind of why we're doing it.
00:37:05.951 --> 00:37:11.981
Which is, which is to say, like we're kind of in the era of Friendster and Facebook hasn't happened.
00:37:11.981 --> 00:37:15.574
Or in the era of, like Shockley Semiconductor and Facebook hasn't happened.
00:37:15.574 --> 00:37:19.887
Or we're in the era where, like, we're still using taxi cabs and Uber hasn't happened.
00:37:19.887 --> 00:37:37.432
In other words, like I think it's like and I really like this point which is like we're doing it because, like, if we can actually succeed, not only can we make our company successful, which is great, but hopefully we can inspire all these other founders and engineers and entrepreneurs and investors and operators, everyone else to do AI and robotics.
00:37:37.432 --> 00:37:40.231
And I was like holy crap, that's pretty cool, that's pretty exciting.
00:37:40.231 --> 00:37:46.596
And so we I went home and I was like, okay, like that's gotta be our mission statement, which is like it's not about chefs succeeding per se.
00:37:46.596 --> 00:37:52.025
It's like how can we kind of inspire other founders to do AI and robotics?
00:37:52.025 --> 00:37:54.771
And it kind of reminded me of this, like we were talking about Olympics at the beginning.
00:37:54.771 --> 00:38:09.619
We were like the four minute mile right Before, nobody really thought about, nobody thought the four minute mile was possible, and then one person did it, and then there's a ton of four minute milers, right, I think that's exactly what I hope to be able to do with Chef, which is the food industry is huge.
00:38:09.619 --> 00:38:11.987
It's arguably one of the biggest industries on planet Earth.
00:38:11.987 --> 00:38:18.168
I hope that we can make Chef into an insanely successful company, deploy tens of thousands, millions of robots Great, amazing.
00:38:18.168 --> 00:38:24.938
But I think more than that, I hope that what we can do is inspire other people to take that and solve these other really important problems.
00:38:24.938 --> 00:38:27.177
And what are those other important problems, right?
00:38:27.177 --> 00:38:30.090
So let's talk about that, these other really important problems, and what are those other important problems, right?
00:38:30.090 --> 00:38:30.753
So let's talk about that.
00:38:30.773 --> 00:38:36.735
One kind of simple way to think about this is kind of literally, go to the Bureau for Labor Statistics and go down and what are the biggest labor forces on planet Earth?
00:38:36.735 --> 00:38:41.474
Right, and on an international scale, it's actually agriculture and food.
00:38:41.474 --> 00:38:47.992
So I mean, outside of the US, agriculture is still number one, actually, from a labor force perspective.
00:38:47.992 --> 00:38:51.516
I think labor force force is a good proxy for the market size for robotics, I would say.
00:38:51.516 --> 00:38:55.715
But in the US you kind of just go down, you're like, okay, what are the biggest occupations?
00:38:55.715 --> 00:39:02.065
And then each of those has literally like opportunities for like trillion dollar companies, literally like it's a billion dollar market.
00:39:02.065 --> 00:39:07.594
It's a trillion dollar markets because the human labor market is $45 trillion of global GDP.
00:39:07.594 --> 00:39:14.079
It's like any particular market is a trillion dollar, not any, but most of the big ones are trillion dollar markets, yeah, yeah.
00:39:14.079 --> 00:39:19.869
So I mean I'm pretty excited about like obviously I'm really excited about food just because I think I'm excited about food, because it's a really big market.
00:39:19.869 --> 00:39:21.994
It's the number one labor shortage in the US.
00:39:21.994 --> 00:39:28.931
It's over a million jobs unfilled and it seems technically tractable, like it seems like it's possible in the short term.
00:39:29.351 --> 00:39:31.012
But I think there's a lot of others that I'm excited about.
00:39:31.012 --> 00:39:34.657
So construction is one that I think actually is getting a little bit of love.
00:39:34.657 --> 00:39:40.202
There's Dusty Robotics, there's Built Robotics, there's some companies there, civ Robotics, there's some love there.
00:39:40.202 --> 00:39:47.547
But I think construction is one of those mega industries where there's so much more that we could do there.
00:39:47.547 --> 00:39:48.829
So I think, and I hope, that there's more there.
00:39:48.889 --> 00:39:55.320
I think medical robotics is one where you know, obviously, with the DaVinci system, with Intuitive Surgical, I think that's actually a pretty crazy company.
00:39:55.320 --> 00:39:56.121
It's worth like 70.
00:39:56.121 --> 00:39:59.394
I didn't check lately, but last I checked it was like $70 billion in market cap.
00:39:59.394 --> 00:40:12.400
But the Intuitive robot, the DaVinci robot, is still like controlled by a physician, right, it's still a surgeon who's actually controlling it and and the robots kind of just helping essentially prevent you from having errors.
00:40:12.400 --> 00:40:16.762
But, but I think more, more surgical robots right, would be actually really, really important.
00:40:17.282 --> 00:40:22.273
And then and then I think I think, like right now there's a lot of like robotics companies that are in manufacturing warehousing.
00:40:22.273 --> 00:40:25.931
I think mainly that's because of the rise of e-commerce as of late and I think that's great.
00:40:25.931 --> 00:40:42.036
I I think, yes, we need, we need that, but then you know, you, you that, but then you just go to manufacturing plants, general manufacturing plants, and there's literally thousands of people a lot more than you might expect who are just doing these manual tasks any from machine tending and just working with different equipment.
00:40:42.036 --> 00:40:52.976
Essentially, there's some tasks that are still manual, like clothing, and stuff is a very manual task clothing manufacturing, making shoes, so anyways, like basically take any job in the physical world.
00:40:52.976 --> 00:40:58.324
There's literally like at least a 10 to 100 billion, potentially a trillion dollar company to be made.
00:40:58.324 --> 00:40:59.347
This is amazing.
00:40:59.688 --> 00:41:00.190
Yeah, yeah.
00:41:00.190 --> 00:41:10.585
So it's like the goal is to sort of like lay the scaffolding or sort of like be the form of them either, but in a way that like, well, here's all this other knowledge, that like, or here's sort of the learnings.
00:41:10.585 --> 00:41:20.594
And in the same way I'm going to make the Amazon comparison it's like hey, they were able to ship on existing freight lines and they had the internet.
00:41:20.594 --> 00:41:26.206
They didn't have to do a bunch of this innovation, just like oh, all these pieces sort of came together, oh, what is now possible?
00:41:26.206 --> 00:41:40.110
And it got people asking the question, I think, when people see 22 million mail served like whoa, oh, this is working, oh, we have to be very specific about the problem we're solving.
00:41:40.231 --> 00:41:41.487
Oh, and building stuff for customers.
00:41:41.487 --> 00:41:42.735
I'm like, okay, what else can we replace?
00:41:42.735 --> 00:41:45.244
People are going to be like, okay, cool, this can be done.
00:41:45.244 --> 00:41:47.068
Chef's doing it in food.
00:41:47.068 --> 00:41:51.125
Okay, well, I'm going to go try and be able to market on shoe construction.
00:41:51.125 --> 00:41:53.650
It's like, sweet, cool, all right, go for it.
00:41:53.650 --> 00:41:55.293
I guess.
00:41:55.293 --> 00:41:57.054
What does our world look like then as a result?
00:41:57.054 --> 00:41:59.175
Like what's our 20 years out?
00:41:59.175 --> 00:42:01.117
Have you seen Wall-E?
00:42:01.117 --> 00:42:02.237
I have seen Wall-E.
00:42:02.237 --> 00:42:06.106
I have some interesting takes on Wall-E, but let's do yours not over it.
00:42:06.920 --> 00:42:12.867
I watched Wall-E and this was like obviously like it's an old movie now, but it was pretty cool, right.
00:42:12.867 --> 00:42:17.264
And one thing that was really interesting is like there are some like general purpose robots in Wall-E.
00:42:17.264 --> 00:42:21.414
There are some general purpose robots in WALL-E, but almost all the robots are for some reason.
00:42:21.414 --> 00:42:27.090
Wall-e is a trash compacting robot, but the trash compacting robot doesn't need to do some other task.
00:42:27.090 --> 00:42:27.432
In our use case.
00:42:27.432 --> 00:42:35.753
This is kind of coming back to humanoids, I guess a little bit, but in our use case, which is assembling these meals, having a humanoid is negative value.
00:42:35.753 --> 00:42:38.284
It's like you don't need that.
00:42:38.284 --> 00:42:40.969
The solution works as good as it needs to.
00:42:40.969 --> 00:42:45.949
Anyways, what I mean to say is like what I'd hope to create is is is a world like WALL-E, right?
00:42:45.949 --> 00:43:39.530
Which is like you have tens of millions of kind of these robots that are kind of doing these day-to-day tasks, and then you have you have some kind of general purpose robots that are doing specific tasks that like make sense for those general purpose robots Like I think the toad carrying application and a lot of like automotive plants actually is a good example where, like great that example, I can see that making sense, and maybe there's millions of those tens of millions of, like, specific robots for some specific thing, and then millions of robots for these more general purpose tasks, and then I think what ideally well, not ideally, but almost certainly what's going to happen, is what's happened in prior kind of revolutions, industrial revolutions, where it's like the tractor or like the steam engine or the printing press or the G-throw toll, like what's always happened is these inventions create, like, basically they reduce the cost and the economy swells.
00:43:39.530 --> 00:43:42.023
And the economy swells Because the economy swells.
00:43:42.023 --> 00:44:07.536
Business people open up more businesses, right, they create more factories and more stores or more whatever, and now there's more jobs and this whole cycle continues and basically the abundance increases, the price of goods decreases, and I think abundance is really the goal, which is how do you like, in the food world, how do you create a world where the cost of fresh food goes down marginally so that it's equivalent to the cost of fast food, for example?
00:44:07.536 --> 00:44:18.969
Or how do you create a world where housing is, because of this construction innovation, is much more affordable, right, it's not affordable housing, but rather all housing can be much cheaper than it needs to.
00:44:18.969 --> 00:44:21.813
I think, really creating a world where the affordable right, like it's not affordable housing, but rather all housing, can be much cheaper than it needs to.
00:44:21.771 --> 00:44:22.903
I think, really creating a world where the abundance right, right, we're like, we're like.
00:44:22.903 --> 00:44:24.248
I don't think there's going to be this world.
00:44:24.248 --> 00:44:27.077
I do not think there's going to be a world where, like, humans are out of work.
00:44:27.077 --> 00:44:36.530
Right, because anytime there's like, like the invention of social media, for example, led people whose whole job it is to do social media.
00:44:36.530 --> 00:44:39.813
That was not a job before, so there's going to be a lot of new jobs created.
00:44:39.813 --> 00:44:50.545
I don't think people are going to be out of a job, but hopefully they can like, invent and they can wander and they can do these things that they previously couldn't do and do things that are better for humans and robots do these mundane tasks.
00:44:51.027 --> 00:44:51.802
Totally Okay.
00:44:51.802 --> 00:45:02.775
I'm going to ask this sort of obvious sort of like challenging do you think about?
00:45:02.775 --> 00:45:08.320
Is sort of someone building this, this technology, like how do we, how do we sort of make sure we don't fall into the?
00:45:08.320 --> 00:45:19.525
Just sort of sit in big like big chairs and, be you know, drink slushies and just plug in tune in, drop out, uh, that a lot of people may find super attractive.
00:45:20.706 --> 00:45:22.208
Yeah, I think it's a good question.
00:45:22.208 --> 00:45:27.653
My opinion is that that's not going to happen so long as we don't kind of screw up.
00:45:27.653 --> 00:45:33.561
From a policy perspective, I would say so I mean what I mean specifically.
00:45:33.561 --> 00:45:44.911
I guess what I'm referring to is like I think, like COVID was actually a really good example of this where, like people gave out all these stymie checks and right, like a lot of people just left work and they kind of stayed home and they invested in, like all these like GameStop and all the games.
00:45:44.931 --> 00:45:46.335
Yeah, degenerate online gambling.
00:45:46.335 --> 00:45:49.949
Yeah, that only happened because there's an abundance of money and people did not have work.
00:45:49.949 --> 00:45:54.885
Right, in this case, I think, like it's not that, it's not that like I.
00:45:54.885 --> 00:46:00.235
I think if we don't do something dumb from a policy perspective, I think there's just going to be people.
00:46:00.235 --> 00:46:05.887
You actually have to retrain yourself, you have to figure out something new, you have to do that work.
00:46:05.887 --> 00:46:08.929
But in our case, for example, nobody's being laid off.
00:46:08.929 --> 00:46:19.277
They're becoming roper operators, they're doing other tasks, they're actually furthering their career, and I think people need money First of all.
00:46:19.277 --> 00:46:21.887
They need money, but also they need to feel like they're making progress in life.
00:46:21.887 --> 00:46:22.949
People want to do something.
00:46:22.949 --> 00:46:35.664
So my opinion is I don't actually think that's going to happen and I think the only thing, the best thing we can do to prevent that just prevent some kind of governmental policy thing that's going to provide an influx of free capital.
00:46:35.664 --> 00:46:38.929
Basically, I think I'm not a fan of UBI and things like that.
00:46:38.929 --> 00:46:41.954
Basically, for example, on that note, yeah, I'm with you on that.
00:46:42.195 --> 00:46:45.336
Like the new study that just came out that Sam Altman funded was like very revealing.
00:46:45.336 --> 00:46:48.581
People were like, oh, hey, actually this doesn't move the needle the way you think it does.
00:46:48.581 --> 00:46:57.349
I mean, the other thing with that is like what we want to do is sort of hey, you don't have to.
00:46:57.349 --> 00:47:00.614
Like labor force has shifted, so there's all these other options for you.
00:47:00.614 --> 00:47:07.809
Now, like you could just watch Netflix and scroll TikTok, but there's a spiritual calling.
00:47:07.809 --> 00:47:09.306
I was like this is my call to adventure.
00:47:09.306 --> 00:47:09.989
Like oh, wait a minute.
00:47:09.989 --> 00:47:11.202
Like is this what I want to do?
00:47:11.202 --> 00:47:12.144
Or like what else can I do?
00:47:12.144 --> 00:47:13.208
I can do anything.
00:47:13.208 --> 00:47:14.911
Anything is possible.
00:47:14.911 --> 00:47:17.074
So let me, I want to go start a robotics company.
00:47:17.074 --> 00:47:26.063
It's like that's where I'm like my sort of like hero's journey, my own sort of adventures, like finding a mountain worth climbing and going to pursue it, and then I think that's right.
00:47:26.103 --> 00:47:28.925
I mean like I think that's exactly right, which is like kind of sick.
00:47:28.925 --> 00:47:34.762
I think the goal of antibodies is to kind of continue going up Maslow's hierarchy of needs, right like towards self flexualization.
00:47:34.762 --> 00:47:41.304
And I think, like I think there's some, there's some would say that that's like not possible, like anytime.
00:47:41.304 --> 00:47:42.746
You have some people that can self-isolate.
00:47:42.746 --> 00:47:48.688
You need to have some people that cannot just like but but, but, I think generally bringing the entire everybody closer too.
00:47:48.688 --> 00:47:50.210
And maybe one interesting example of this is like gig economy.
00:47:50.210 --> 00:48:13.150
So when I talk to people in the food industry, for example, they explicitly say that it seems like things got a little bit worse in this decade or sorry, not this decade, the 2010s and they kind of credit that to the gig economy.
00:48:13.530 --> 00:48:19.262
Which is to say, previously, if you were at a college, right out of college or in high school, whatever, what do you do?
00:48:19.262 --> 00:48:20.782
You go to a restaurant or something right.
00:48:20.782 --> 00:48:22.043
You go to work at a restaurant.
00:48:22.043 --> 00:48:23.445
But now what can you do?
00:48:23.445 --> 00:48:42.456
Instead of being all sweaty in a hot kitchen working on somebody else's hours, you go work for Uber and you're kind of sitting in an AC car and it's much more nice and things like that and that was only possible because of technology, and I would hope that, to your point, as more and more technology happens and kind of gets productionized, you usually have more choice.
00:48:42.456 --> 00:48:45.282
As opposed to having to do it, you can do it.
00:48:46.063 --> 00:48:53.228
Yeah, and I think there's sort of a cultural component to it too, which is like making certain things more desirable than others.
00:48:53.228 --> 00:49:09.565
It's like, hey, it's a good thing to go, be uncomfortable for a while, but like hey, there's so much choice in the world, like you don't have to do this forever, but like this is you know you, like you level up, it's like an opportunity for you to improve and get better, rather than like just be comfortable 100% of the time.
00:49:10.688 --> 00:49:26.253
Yeah, I'm very excited about this feature, right, like I think like I've been very lucky in my personal life where I've been able to focus on things that I'd like to do, right, and I think that's amazing and it seems like same with you, but that's for so few people.
00:49:26.253 --> 00:49:33.896
I mean, it's literally like the vast majority of the world, billions of people right, they're kind of working paycheck to paycheck and they're just doing it because they have to do it.
00:49:33.896 --> 00:49:36.786
But wouldn't it be incredible if they had a little more choice?
00:49:36.786 --> 00:49:42.146
Right, like some of the basic goods of life food, shelter, all these things are cheaper.
00:49:42.146 --> 00:49:44.481
Right, there's more abundance when it comes to those materials.
00:49:45.242 --> 00:49:49.893
And it's interesting, like economics is a study of supply and demand.
00:49:49.893 --> 00:49:55.422
Right, it's a study of kind of it's kind of the study of scarcity, if you will, is the way one like economist put it.
00:49:55.422 --> 00:50:09.820
So, if you don't have scarcity, if you have abundance, then a lot of economics and it breaks down because you already have a lot of these goods just like freely available, right, and and then hopefully these people can kind of do do whatever they would like to do.
00:50:09.820 --> 00:50:14.043
Now, that's not exactly whatever they would like to do, is probably still a job, but they have more choice in that job.
00:50:14.324 --> 00:50:22.490
I'm going to save the discussion of like okay, what's sort of like a abundant future, sort of like economic system.
00:50:22.490 --> 00:50:23.293
Look like that's you know post.
00:50:23.293 --> 00:50:26.139
That's not like capitalism or communism or anything else.
00:50:26.139 --> 00:50:26.943
Like what is a new.
00:50:26.943 --> 00:50:34.251
What is this sort of like new economic system look like when we have incredible abundance through you know, limitless energy and material goods?
00:50:34.251 --> 00:50:35.945
And like automated labor and AI.
00:50:35.945 --> 00:50:38.286
But that is, that is for another day.
00:50:38.286 --> 00:50:51.367
One question before we wrap other than parts of Wall-E where else have you seen a positive vision for a future enabled by robotics published Like in books, films, essays?
00:50:51.367 --> 00:50:54.929
When you sort of are using what do you point to?
00:50:55.300 --> 00:51:08.086
This is sort of like what we're going for I mean, a few thoughts are coming to mind and some of them are like anti-stories, right, so, like isaac asmo writes a lot of like, not anti-stories, but like he kind of talks about like, like with irobot and others.
00:51:08.086 --> 00:51:09.530
Right, what can go wrong?
00:51:09.530 --> 00:51:15.443
Yeah, right, I mean, I think the classic example that many people might call out is like things like star wars right, we have these.
00:51:15.443 --> 00:51:25.666
Like robots are kind of like these little you know beings, that kind of help you, I, I would say that, like you know, honestly, like the thought that's also going to my hands like there hasn't been like a very clear.
00:51:25.666 --> 00:51:30.565
There isn't like the story of like wow, like that's so, like, which is kind of like the whole reason.
00:51:30.585 --> 00:51:34.643
I kind of came back to this, which is like I was thinking about quitting and it's like okay, like, why are we doing this?
00:51:34.643 --> 00:51:37.246
It's like there wasn't a clear like aha, that's why we're doing it.
00:51:37.246 --> 00:51:56.304
I think, like, so from like, I guess I'll just say for me personally right, like, who inspires me is like people who have like, who have like done it, and and I would say like, I think, like you know, elon is somebody who's like inspired probably billions of people on this planet because he, like he took this thing that like nobody thought was possible and he made it doable.
00:51:56.304 --> 00:52:06.489
And you know, you look at his like background, it's like sure, like yeah, there's some all these theories about like whatever, but like I'm sure he's a smart person, but like okay, if you can do it, why can't any of us?
00:52:06.489 --> 00:52:13.231
Right me personally, I'm less inspired by like science fiction or like fiction or writing or any of that.
00:52:13.271 --> 00:52:21.793
I think more inspired by like biographies, if you will, like people who have like done it themselves, and I think in a lot of industries there are success stories, right.
00:52:21.793 --> 00:52:27.391
So, like, if I'm building a marketplace, I'm going to be inspired by Travis Kalanick or Tony Zhu.
00:52:27.391 --> 00:52:30.242
Some people have opinions, but whatever, maybe Tony Zhu, you're inspired by whatever.
00:52:30.242 --> 00:52:39.833
If I'm running a semiconductor company, maybe I look towards, like you know, the early guys at Intel and Andy Grove and maybe look at Morris Chang at TSMC.
00:52:39.833 --> 00:52:42.485
But who do I look for for AI and robotics?
00:52:42.485 --> 00:52:48.708
Well, I can maybe look at like I don't know, like it's kind of something we talked about, but there isn't that company or that person.
00:52:48.708 --> 00:52:50.021
So I hope that.
00:52:50.021 --> 00:53:00.340
I hope and this is a very bold claim but I hope that I and my peers in robotics and help like really become one of the early success stories and help create that playbook for others.
00:53:00.340 --> 00:53:01.664
Amen.
00:53:01.824 --> 00:53:02.105
Love it.
00:53:02.105 --> 00:53:04.088
That's a great note to wrap on.
00:53:04.088 --> 00:53:05.913
Where can people find you?
00:53:05.913 --> 00:53:07.568
Where do you want to point them to?
00:53:07.789 --> 00:53:09.018
I think Twitter and LinkedIn are the best.
00:53:09.018 --> 00:53:20.813
I'm just at Rajat Pagiriya, find me both there and then If you are interested you know if you are, you know, interested in Chef system, chefroboticsai, and, on the other hand, if you think this is an exciting mission, also feel free to reach out as well.
00:53:21.253 --> 00:53:21.574
Amazing.
00:53:21.574 --> 00:53:23.344
Well, man, this has been a blast.
00:53:23.344 --> 00:53:24.186
Thanks so much for coming on.
00:53:24.186 --> 00:53:25.570
Yeah, thank you, Cam.
00:53:25.570 --> 00:53:25.931
This was fun.
00:53:25.931 --> 00:53:29.846
Thanks for joining us for this episode of the Build the Future podcast.
00:53:29.846 --> 00:53:32.052
If you loved it, we'd be really grateful if you share it with a friend.